Asset Allocation in a

Value-at-Risk Framework

Ronald Huisman, Kees G. Koedijk and Rachel A.J. Pownall

April 1999

In this paper we develop an asset allocation model which allocates assets by maximising
expected return subject to the constraint that the expected maximum loss should meet
the Value-at-Risk limits set by the risk manager. Similar to the mean-variance approach a
performance index like the Sharpe index is constructed. Furthermore it is shown that the
model nests the mean-variance approach in case of normally distributed expected
returns. We provide an empirical analysis using two assets: US stocks and bonds. The
results highlight the influence of non-normal characteristics of the expected return
distribution on the optimal asset allocation.

Correspondence:
Rachel A. J. Pownall
Erasmus University Rotterdam, Faculty of Business Administration
Financial Management, 3000 DR Rotterdam
The Netherlands
Tel: +31 10 4081255 Fax: +31 10 4089017

All authors are at the Erasmus University in Rotterdam. Koedijk is also at Maastricht University, and
CEPR. The respective email addresses are rhuisman@fac.fbk.eur.nl, ckoedijk@fac.fbk.eur.nl, and
rpownall@fac.fbk.eur.nl. All errors pertain to the authors. The authors would like to thank Frans de
Rhoon and participants at the Rotterdam Institute for Financial Management lunch seminar series for their
comments.
1 Introduction

Modern portfolio theory aims to allocate assets by maximising the expected risk

premium per unit of risk. In a mean-variance framework risk is defined in terms of the

possible variation of expected portfolio returns. The focus on standard deviation as the

appropriate measure for risk implies that investors weigh the probability of negative

returns equally against positive returns. However it is highly unlikely that the perception

of investors to downside risk faced on investments is the same as the perception to the

upward potential. Indeed risk measures such as semi-variance were originally constructed

in order to measure the downside risk separately.

Another approach that has been taken is to incorporate downside risk directly into the

asset allocation model. The optimal portfolio is then selected by maximising the expected

return over candidate portfolios so that some shortfall criterium is met. Leibowitz and

Kogelman (1991), and Lucas and Klaassen (1998) for example construct portfolios by

maximising expected return subject to a shortfall constraint, defined such, that a

minimum return should be gained over a given time horizon for a given confidence level.

Roy (1952) and Arzac and Bawa (1977) define the shortfall constraint such that the

probability that the value of the portfolio falling below a specified disaster level is limited

to a specified disaster probability; these are examples of the safety-first approach to asset

allocation. Although the concept of using shortfall constraints is more in line with

investors’ perception to risk, their applicability is rather limited since the disaster levels,

minimum returns, confidence levels or disaster probabilities are hard to specify.

In this paper we extend the literature on asset allocation subject to shortfall constraints.

We adress the criticism concerning the definition of disaster levels and probabilities

2
through the use of Value-at-Risk (VaR). Banks and financial institutions have adopted

VaR as the measure for market risk1, whereby VaR is defined as the maximum expected

loss on an investment over a specified horizon given some confidence level2. It therefore

reflects the potential downside risk faced on investments in terms of nominal losses. We

therefore show how VaR can be implemented into the asset allocation framework using

shortfall constraints. The advantage being that the shortfall constraint is then clearly

defined in terms of a widely accepted market risk measure. Furthermore the

methodology extends the richness of VaR as a risk management tool. At present risk

managers determine the VaR of their portfolios, for a chosen confidence level, whereby

VaR is used as an ex-post measure to evaluate the current exposure to market risk and the

decision as to whether exposure ought to be reduced. The asset allocation framework

subject to a VaR shortfall constraint extends the importance of VaR as an ex-ante market

risk control measure. In our framework risk managers thus define a VaR limit, i.e. set the

maximum allowed loss that will only be exceeded by a small probability, and then let the

portfolio manager allocate assets in such a way that the maximum expected loss on the

portfolio meets the VaR limit.

In developing the model we proceed along the lines of Arzac and Bawa (1977). Their

approach is superior to the others mentioned above, since they are able to derive a

performance measure equivalent to the Sharpe index, which gives portfolio managers an

easy tool with which they can evaluate the efficiency of several (candidate) portfolios.

Furthermore market equilibrium can be derived under specific assumptions regarding the

distributional characteristics of expected returns on portfolios. For example the asset

allocation model leads to the same optimal portfolio as a mean-variance approach in the

case of normally distributed expected returns. It therefore generates the comfortable

feeling that the framework is closely related to traditional models. It is widely known

3
however that the expected returns on many financial assets are not normally distributed

(they are not symmetric and exhibit excess leptokurtosis); we are also able to incorporate

this issue into the asset allocation decision. The model need not make any assumptions

regarding the skewness and the structure of the tails of the distribution, thus enabling us

to choose a distribution, which can correctly reflect the apparent tail fatness of financial

return distribution: an important feature for a framework that focuses on downside risk

when finding the optimal allocation of assets.

The plan of the paper is as follows. We introduce the framework in the following section.

The third section then provides empirical results of the optimal portfolio allocation for a

variety of asset classes. We also shall address the importance of the non-normal

characteristics of expected return distributions in such a framework. Conclusions and

practical implications are drawn in the final section.

2 Methodology

In this section we present a portfolio construction model subject to a VaR limit set by

the risk manager for a specified horizon. In other words we derive an optimal portfolio

such that the maximum expected loss would not exceed the VaR for a chosen

investment horizon at a given confidence level. Using VaR as the measure for risk in this

framework is in accordance with the banking regulations in practice and provides a clear

interpretation of investors’ behaviour of minimising downside risk. The degree of risk

aversion is set according to the VaR limit; hence avoiding the limitations of expected

utility theory as to the degree of risk aversion, which an investor is thought to exhibit.

4
2.1 The asset allocation problem and shortfall constraints

Suppose the risk management department sets the portfolio manager’s VaR limit on an

amount W(0) to be invested for an investment horizon T. The portfolio manager can

invest this amount plus an amount B representing borrowing (B > 0) or lending (B < 0).

The manager may invest in n assets whereby γ(i) denotes the fraction invested in the

risky asset, i. Hence the γ(i)’s must sum to one. Let P(i,t) be the price of asset i at time t (t

= 0 reflects the current decision period). Equation (1) gives the initial value of the

portfolio as the following budget constraint:

n
W ( 0 ) + B = ∑ γ ( i ) P ( i ,0 ) (1)
i =1

According to 0 the manager needs to choose the fractions γ(i) to be invested with wealth

W(0) and the amount to be borrowed or lent at time 0: we assume rf is the interest rate at

which the investor can borrow and lend for the period T. The portfolio allocation

problem arises on allocating the assets in the portfolio and choosing the amount to

borrow or lend such that the maximum expected level of final wealth is achieved.

The introduction of Value-at-Risk into risk management requires the portfolio manager

to be highly concerned about the value of the portfolio falling below the VaR constraint.

VaR is defined as the worst expected loss over a chosen time horizon within a given

confidence interval c (see Jorion 1997). For example a 99% VaR for a 10-day holding

period3, implies that the maximum loss incurred over the next 10 days should only

exceed the VaR limit once in every 100 cases. Chosing the desired level of Value-at-Risk

as VaR* we therefore formulate the downside risk constraint as follows:

Pr{W ( 0 ) − W ( T ) ≥ VaR *} ≤ (1 − c ) (2)

5
Pr denotes the expected probability conditioned on the information available at time 0.

Equation 0 is equivalent to:

Pr{W ( T ) ≤ W ( 0 ) −VaR *} ≤ (1 − c ) (3)

VaR is thus the worst expected loss over the investment horizon T, that can be expected

with confidence level c. The investor’s level of risk aversion is reflected in the VaR level,

and the confidence level associated with it. Hence the optimal portfolio which is derived

such that equation 0 holds will reflect this.

2.2 Optimal portfolio construction

Portfolio construction subject to a shortfall constraint is not new. Leibowitz and

Kogelman (1991) and recently Lucas and Klaassen (1998) construct portfolios by

maximising expected return subject to a shortfall constraint defined such that a minimum

return should be gained over the given time horizon within a given confidence level.

Arzac and Bawa (1977) developed a similar framework according to the safety-first

principle introduced by Roy (1952). In that framework investors maximise the expected

return of their portfolio subject to a shortfall constraint defined such that the probability

that the end-of-investment horizon value of the portfolio falls below a disaster level s is

smaller than a disaster probability α. Arzac and Bawa develop the optimal asset

allocation for risk-averse investors and are also able to show that the model nests the

CAPM for specific distribution characteristics of the expected returns. We therefore

believe that the Arzac and Bawa model is the most serious contender to a theory of

portfolio choice for risk-averse investors. The limitation however is in specifying the

investor’s disaster level and the associated probability of disaster.

6
The introduction of VaR however provides us with a shortfall constraint (denoted by

equation 0) that fits perfectly into the Arzac and Bawa framework. We therefore build

upon their results to derive an optimal asset allocation model. The investor is interested

in maximising wealth at the end of the investment horizon. Let r(p) be the expected total

return on a portfolio p in period T; assume that asset i is included with fraction γ(i,p) in

portfolio p. The expected wealth from investing in portfolio p at the end of the

investment horizon becomes:

E0 (W (T , p )) = (W ( 0 ) + B )(1 + r ( p )) − B(1 + r f ) (4)

Substituting in for B as given in equation (1), we are able to express final wealth in terms

of the risk-free rate of return and the expected portfolio risk premium (r(p)-rf):

n
E0 (W (T , p )) = W ( 0 )(1 + r f ) + ( ∑ γ ( i , p )P ( i ,0 ))( r ( p ) − r f ) (5)
i =1

Equation 0 shows that as long as the expected risk premium is positive a risk-averse

investor will always invest some fraction of his wealth in the risky assets. In order to

determine the optimal portfolio that maximises the expected final wealth subject to the

VaR constraint 0 we substitute 0 into 0:

VaR * +W ( 0 )r f
Pr{r ( p ) ≤ r f − n
} ≤ (1 − c ) (6)
∑ γ ( i , p )P ( i ,0 )
i =1

Equation 0 simply defines the quantile q(c,p) that corresponds to probability (1-c) that

can be read off the cdf of the expected return distribution for portfolio p. This quantile is

used to derive the following expression from 0:

n VaR * +W ( 0 )r f
∑ γ ( i , p )P ( i ,0 ) = r f − q( c , p )
(7)
i =1

7
Substituting 0 back into 0 leads us to the following expression for the expected final

wealth in terms of the quantile q(c,p):

(r ( p ) − r f )
E0 (W (T , p )) = W ( 0 )(1 + r f ) + (VaR * +W ( 0 )r f ) (8)
( r f − q( c , p ))

Dividing 0 by initial wealth W(0) we obtain the following expression for the expected

return on the initial wealth:

W (T , p ) (r ( p ) − r f )
E0 ( ) = (1 + r f ) + (VaR * +W ( 0 )r f ) (9)
W ( 0) (W ( 0 )r f − W ( 0 )q( c , p ))

It can be seen from equation 0 that the final expected return on wealth is maximised for

an investor concerned about the downside risk for the portfolio which maximises S(p) in

equation (10). We denote this maximising portfolio as p′.

r( p) − rf
p ' : max S ( p ) = (10)
p W ( 0 )r f − W ( 0 )q( c , p )

Note that although initial wealth is in the denomenator of S(p) it does not affect the

choice of the optimal portfolio since it is only a scale constant in the maximisation. The

asset allocation process is thus independent of wealth. The advantage however of having

initial wealth in the denomenator of 0 is in its interpretation. S(p) equals the ratio of the

expected risk premium offered on portfolio p to the risk, reflected by the maximum

expected loss on portfolio p that is incurred with probability 1-c relative to the risk-free

rate. Since the negative quantile of the return distribution multiplied by the initial wealth

is the Value-at-Risk associated with the portfolio for a chosen confidence level, then we

are able to derive an expression for the risk faced by the investor as ϕ. Letting VaR(c,p)

denote portfolio p’s Value-at-Risk, the denominator of (10) may be written as:

8
 ϕ ( c , p ) = W ( 0 )r f − VaR ( c , p ) (11)

Such a measure for risk is in fitting with investors’ behaviour of focussing on the risk

free rate of return as the benchmark return with risk being measured as the potential for

losses to be made with respect to the riskfree rate as the point of reference. Indeed the

measure for risk can be seen as a possible measure for regret, since it measures the

potential opportunity loss of investing in risky assets. Investors will therefore only accept

greater returns if they can tolerate the regret occuring from the greater potential wealth-

at-risk. The risk-return ratio S(p), which is maximised for the optimal portfolio p′ can

therefore be written as:

r( p) − rf
p ' : max S( p ) = (12)
p ϕ( c , p )

S(p) is thus a performance measure like the Sharpe index that can be used to evaluate the

efficiency of portfolios (see Sharpe 1994 for more details). Indeed under the assumption

that expected portfolio returns are normally distributed, and the risk free rate is zero, S(p)

collapses to a multiple of the Sharpe index. In this case the VaR is expressed as a

multiple of the standard deviation of the expected returns so that the point at which both

performance indices are maximised will lead to the same optimal portfolio being chosen.

Only a minimal difference in the optimal portfolio weights occurs for positive risk free

rates, so that both approaches will lead to almost identical results. Since our performance

index S(p) does not rely on any distributional assumptions it has the advantage of being

able to incorporate non-normalities into the asset allocation problem, through the use of

other distributional assumptions. The existence of non-normalities may lead to the

choice of different optimal portfolios, an empirical investigation of which we shall

encounter later.

9
The optimal portfolio that maximises S(p) in 0 is chosen independently from the level of

initial wealth. It is also independent from the desired VaR, since the risk measure ϕ for

the various portfolios depends on the estimated portfolio VaR rather than the desired

Value-at-Risk. Investors first allocate the risky assets and then, the amount of borrowing

or lending will reflect by how much the VaR of the portfolio differs from the VaR limit

set; thus two-fund separation holds like in the mean-variance framework. However since

the investors’ degree of risk aversion is captured by the chosen Value-at-Risk level, the

amount of borrowing or lending required to meet the VaR constraint may be

determined. The amount of borrowing is therefore found by substituting (1) and (11)

into equation (7):

W ( 0 )(VaR * −VaR ( c , p ' ))

B= (13)
ϕ '( c , p' )

Of course the critical assumption in finding the maximising portfolio and hence the

optimal asset allocation is in the choice of distributional assumption for the future

distribution of returns. It is to this question which we now turn, with the use of an

empirical example for US stocks and bonds.

3 Optimal Asset Allocation for US stocks and Bonds

In this section we provide an empirical example in which a portfolio manager needs to

select the optimal allocation of US stocks and Bonds such that a VaR constraint is met.

We employ daily data from US stock and bond indices from January 1980 until

December 1998, providing us with 2364 observations. We use data obtained from

Datastream for the S&P 500 Composite Return Index for the US, the 10- Year

Datastream Benchmark US Government Bond Return Index and the 3- Month US

10
Treasury Bill rate for the risk free rate. The average annual return on the S&P 500 over

the sample period was 16.81%, just over twice as high as the average annual return on

the 10-year Government Bond Index of 8.35%. The annual standard deviation is also

higher on the S&P 500 at 13.42% per annum, compared to the less volatile nature of the

Government Bonds with an annual standard deviation of only 6.31%. Both series exhibit

significant negative skewness, -0.45 and –0.42 for the S&P 500 Index and Bond Index

respectively, and significant kurtosis, with greater excess kurtosis on the S&P 500 Index,

6.88, than on the Bond Index, 3.33.

3.1 Optimal allocation using the empirical distribution

To maximise the performance index S(p) in 0 we estimate the expected return r(p) and

the daily Value-at-Risk(c,p) for various combinations of US stocks and bonds using the

whole sample period. Figure 1 shows the efficient VaR frontier for a 95% confidence

level. The VaR level for the 95% level is directly read off from the empirical return

distribution for the various combinations of stocks and bonds.

Insert Figure 1

The VaR efficient frontier is similar to a mean-variance frontier except for the definition

of risk: VaR relative to the benchmark return (ϕ) instead of standard deviation (σ). The

lower-left point represents a portfolio containing 100% bonds and the upper-right

portfolio depicts a 100% investment into stocks.

Below we assume that the portfolio manager needs to select a portfolio that has the same

95% VaR as has been observed in the past. The optimum position on the efficient VaR

frontier, for an investor, who wants to be 95% confident that his wealth will not drop by

more than the daily VaR limit, occurs where the return per unit of risk is maximised. This

11
occurs where equation (9) is maximised, and is independent of the level of wealth. Using

the empirical sample we are therefore able to determine the optimal allocation between

stocks and bonds, whereby the last available 3-month Treasury Bill rate in the sample

period is used for the risk free rate of return. Taking 4.47% as the risk free rate, we find

that the optimal allocation between US stocks and bonds occurs when 36% of wealth is

held in stocks and 64% in bonds for an investor with a VaR limit at the 95% confidence

level. If the desired daily Value-at-Risk, which the risk manager sets is different from the

VaR associated with this 95% empirical estimate, then part of the wealth will need to be

lent, or additional funds borrowed, at the risk free rate in accordance with equation (13).

This involves moving along the Capital Market Line, also shown in Figure 1, until the

desired trade off is attained.

The combinations for stocks and bonds for a variety of confidence levels are provided in

the first two columns of Table 1.

Insert Table 1

Taking the desired VaR level as the daily VaR at the 95% confidence level from the

empirical distribution as our benchmark we can observe how additional lending is

required so that this benchmark VaR level is met for higher confidence levels. We can

therefore see how sensitive the asset allocation decision is to changes in the confidence

level associated with the Value-at-Risk limit. Allocating 36% in Stocks and 64% in Bonds

generates a 95% VaR on the portfolio of $6.86 and of course no borrowing or lending is

required to meet the VaR with 95% confidence. If however the risk manager desires

greater confidence in the probability that the initial wealth will not drop by more than the

VaR level, then the VaR associated with the portfolio allocation will be greater than the

VaR limit and hence results in too much risk being taken4. In order to meet the

12
benchmark VaR less risk will have to be taken and hence $313.20 of the initial $1000

wealth is lent at the risk free rate. In the last column of the table we show the nominal

amounts needed to be lent so that the daily VaR of $6.86 is met for the various

confidence levels. The final portfolio allocations are given in the bottom segment of

Table 1, and we see that the greater the confidence level, hence the lower the risk

tolerance of the investor, the greater the proportion of wealth that needs to be lent at the

risk free rate5.

We have now shown how we can find the optimal allocation from the use of the

empirical distribution, whereby we assumed that future returns are distributed in exactly

the same manner as in the past. The reliance on a large sample is therefore crucial, so

that the quantiles are estimated accurately. The choice of sample period is also extremely

crucial to the estimation process. It may therefore be more desirable to assume a

parametric distribution for characterising the distribution of future returns. Modifying

the methodology in such a way certainly levies some benefits. Firstly, estimation risk is

reduced - especially crucial for high confidence levels - hence providing more accurate

estimates of quantiles. Secondly, quantile estimation is simplified since the quantiles are a

function of the parameters determining the distribution. The characteristic parameters

are derived from the historical data and the quantile is obtained by inferring the quantile

point from the fitted distribution function. Finally, certain parametric assumptions enable

a unique solution to be found for the maximising portfolio, regardless of the quantile

used in the maximisation process. Regardless therefore of the investors risk preferences

the choice of risky assets will be identical up to a scale constant for all investors. Specific

parametric distributional assumptions therefore enables a market equilibrium model to

be developed and hence allows its applicability to be tested empirically by means of a

testable model of asset valuation.

13
If we assume that the future distribution of returns can be accurately proxied by the

normal distribution, the only risk factor in our downside risk measure is the standard

deviation of the distribution. This means that the quantile estimate is merely a multiple of

standard deviations, and our risk measure ϕ in equation (12) is a multiple of the standard

deviation alone. This results in the risk-return trade off being identical to that derived

under the mean-variance framework. The maximisation will occur at the same point as

which the Sharpe ratio is maximised, and the market model will collapse to the CAPM.

Of course since we also have the possibility of assuming different distributional

assumptions, we need not constrain ourselves to optimising our portfolio according to

the first two moments of the distribution only and hence are able to include the

possibilities of non-normalities into asset allocation. Indeed any parametric distribution

used to represent the future distribution of returns, which can provide a unique solution

to our maximising equation, equation (9), regardless of the quantile chosen, will allow for

the derivation of a market equilibrium model.

Before entering the discussion on alternative market equilibrium models it is important

to first analyse the extent to which non-normalities effect the optimal allocation of assets.

We therefore compare the optimal allocation of assets derived using both the normal

distribution and a fatter tailed distribution, the student-t, whereby we use the same

sample period of data as before.

3.2 Optimal Allocation of US Stocks and Bonds under Parametric distributions

We first shall assume that the distribution of future returns can be captured using the

normal distribution. To see how accurate the use of standard deviation alone is as the

measure for risk we can plot the efficient VaR frontier for the risk-return trade off at the

14
95% confidence level for the normal distribution against the empirical VaR frontier

derived earlier. It can also be seen in Figure 1 that at the 95% VaR level the assumption

of normality reflects the actual risk-return trade off fairly well. On average the

assumption of normality for the future distribution of returns at the 95% level means

that the risk is only slightly overestimated6 for a given level of return. The risk is

minimised at the optimal allocation of 40% stocks and 60% bonds, and as mentioned

earlier this optimum will occur at the same point for various confidence levels. This is

presented in table 27.

Insert Table 2

In order to compare how well the normal distribution proxies for the empirical

distribution, we again use the historical VaR at the 95% level as our benchmark VaR

level. With wealth equal to $1000, an amount of $44.43 needs to be lent at the risk free

rate to meet the empirical VaR at the 95% level.

We have seen that the assumption of normality renders the investors attitude to risk as

unimportant in the optimisation process. Only after the optimising portfolio has been

found do the individuals risk preferences come into play. The non-parametric nature of

the empirical distribution however, led to the changing optimum allocation of assets for

various confidence levels8, whereby the optimal allocation of assets resulted in a

proportionally greater increase in lending to meet the desired VaR level for higher

confidence levels. Under the assumption of normality with standard deviation alone as

the measure for risk this effect is not captured, and it appears that for a desired

confidence level of 99%, too aggressive an investment strategy will result. It would

appear that the trade off between risk and return is underestimated for higher confidence

levels. We can see this graphically by comparing the efficient VaR frontiers at the 99%

15
confidence level for both the normal and the empirical distributions. This is presented in

Figure 2.

Insert Figure 2

Risk, as measured by the empirical VaR for the portfolio is higher for all combinations of

stocks and bonds than captured by the use of standard deviation alone. This

underestimation of risk far out in the tails of the distribution will be greater the greater

the deviation from normality9. The greater probability of extreme negative returns in the

empirical distribution implies greater downside risk than is captured by the measure of

standard deviation alone. The use therefore of the normal distribution to assess the risk-

return trade off will result in an incorrect allocation of assets for investors with low risk

tolerance and risk managers wishing to set 99% confidence levels. Since non-normalities

will cause errors to be made in the asset proportions held, it would appear more desirable

to use a parametric distribution that more accurately reflects the distributional

characteristics of financial assets for the whole range of confidence levels associated with

the left tail of the distribution.

The student-t with 5 degrees of freedom, a fatter tailed distribution than the normal

distribution, has been shown to represent the tail of the distribution of many financial

assets more accurately than the normal distribution. In order to determine the extent to

which non-normalities effect the optimal allocation of assets we compare both the

efficient VaR frontier and the optimal allocation of assets using the assumption that

returns are distributed as a student-t with 5 degrees of freedom10. We therefore derive

efficient VaR frontiers using the empirical distribution, against the normal and the

student-t. In figure 1 we see that the student-t with 5 degrees of freedom represents the

empirical trade off at the 95% confidence level fairly well. This confirms that the

16
additional downside risk associated with the presence of fat tails in the distribution only

becomes apparent for investors and risk managers wishing to know with greater certainty

the probability of exceeding their VaR limits.

As we move to higher confidence levels, as shown in figure 2 for the 99% level, we find

that it becomes vital that the additional downside risk from fat tails is incorporated into

the risk-return trade off11. Comparing the optimal portfolio allocation in Table 2 we find

that the proportions held in the various risky assets are again identical to the outcome

when we assumed normally distributed returns. The measure for risk is however greater

when using the student-t distribution and hence the amount of lending required to meet

the same VaR level as before is greater. This is presented in the final two columns of

Table 2. Indeed the proportion of wealth held in the various assets is much more in line

with the optimal allocation when using the empirical distribution.

The use therefore of a risk measure, which is able to capture higher moments of the

distribution appears to capture the true trade-off between risk and return as observed in

financial markets. It indeed may provide a better explanation for the size of the risk

premium, an interesting area for future research. Furthermore the results show that if a

risk manager is concerned about his 99% VaR (as recommended by the regulatory

framework of the Basle Committee) the use of the standard deviation alone in

determining the correct risk-return trade off with which to optimise the allocation

decision is incorrect. The use therefore of a parametric distribution, which is able to

capture some of the additional downside risk, such as the student-t, would appear to

allocate assets more in accord with the risk-return trade off observed in US financial

markets.

17
We find that the use of a student-t distribution with 5 degrees of freedom is better able

than the assumption of normality to correctly assess the risk return trade off in financial

markets, however the exact nature of the parametric distribution which best represents

the distribution of future returns in the negative tail of the distribution is still open to

empirical research. Of further vital importance is the correct assessment of the

correlation between assets as we move further into the tails of the distribution, and the

notion that correlation increases for large negative returns may be able to explain the

under assessment of the downside risk at high confidence levels12.

4 Concluding Remarks

The move towards greater risk management has highlighted the need to be able to

control and manage financial risk. Until now Value-at-Risk as a risk management tool has

been used in the assessment of risk ex-post in financial markets. In this paper we highlight

the need for taking an ex-ante approach to risk management, such that assets are allocated

so that expected return is maximised and that the risk of the portfolio’s value falling

below a critical level is known. The introduction of VaR allows us to develop such an

asset allocation model within a Value-at-Risk framework, so that the focus is on

downside risk rather than standard deviation as the crucial measure for risk. The use of

Value-at-Risk therefore enables the level of risk aversion to be measured, and hence the

desired risk-returns profile of the investor to be attained: a limitation of modern

portfolio theory.

In this paper we provide a model for asset allocation which is able to move away from

the use of standard deviation alone as the approriate measure for risk in financial

markets. We focus on the use of downside risk, as measured by Value-at-Risk, and hence

18
are able to allocate assets more in accordance with the risk perceptions which investors

hold. The model is derived without having to impose distributional assumptions about

the future distribution of returns, and hence enables an approach which is also able to

encompasses the mean-variance approach, however through distributional assumptions

other than normality we are able to incorporate non-normalities into the asset allocation

process. The use of the normal distribution, or the student-t, allows for a market

equilibrium model to be derived13, whereas the use of normality enables the model to

collapse to the CAPM14.

Empirical results for US stock and bonds provide evidence of additional downside risk

from skewness and kurtosis, and thus the use of the normal distribution results in an

incorrect estimation of the risk-return trade off for investors wanting to know the

probability of their portfolio falling below the VaR level with high confidence. Since the

Basle committee currently recommends bank’s capital requirements to be a multiple of

the 99% Value-at-Risk, allocating assets such that the 99% VaR level is met is of crucial

importance in practice. Desirable is therefore a model of asset allocation, which is well

able to assess the risk-return trade-off observed in financial markets, such that the VaR

constraint is met. The inclusion therefore of a non-normal distribution in asset allocation

models would appear to be much more desirable than the present assumption that assets

are normally distributed in the tails. A move away from current portfolio theory, such

that assets can be allocated to meet risk management constraints, and the possible

inclusion of non-normalities in the asset allocation process as outlined in this paper

therefore appears a crucial direction in which to move.

19
5 References

Arzac, E.R. and V.S. Bawa (1977) ‘Portfolio Choice and Equilibrium in Capital Markets

with Safety-First Investors’. Journal of Financial Economics, 4, 277-288.

Harlow, W. V. (1991) ‘Asset Allocation in a Downside-Risk Framework’, Financial

Analysts Journal, 28.

Huisman, R., C.G. Koedijk, and R.A. Pownall (1998) ‘VaR-x: Fat Tails in Financial Risk

Management’. Journal of Risk, 1, 47-61.

Jansen, D.W., C.G. Koedijk, and C.G. de Vries (1998) ‘Portfolio Selection with Limited

Downside Risk’. LIFE Working paper.

Jorion, P. (1997) Value at Risk: The New Benchmark for Controlling Derivatives Risk, McGraw-

Hill Publishers.

Leibowitz, M.L. and S. Kogelman (1991) ‘Asset Allocation under Shortfall Constraints’.

Journal of Portfolio Management, Winter, 18-23.

Longin, F. and B. Solnik (1995) ‘Is the International Correlation of Equity Returns

Constant: 1960-1990?’. Journal of International Money and Finance, February 1995.

Lucas, A. and P. Klaassen (1998) ‘Extreme Returns, Downside Risk, and Optimal Asset

Allocation’. Journal of Portfolio Management, Fall, 71-79.

20
Roy, A.D. (1952) ‘Safety-First and the Holding of Assets’. Econometrica, 20, 431-449.

Sharpe, W. (1994) ‘The Sharpe Ratio’. Journal of Portfolio Management, 21, 49-58.

Stutzer, M. (1998) ‘A Portfolio Performance Index and Its Implications’. U. of Iowa

Working Paper.

21
Table 1

Optimal Asset Allocation using Empirical Distribution

The table gives the optimal allocation between Stocks and Bonds using data on the S&P 500 Composite

Returns Index and the 10- Year Datastream US Benchmark Government Bond Index over the period

January 1980 - December 1998. The risk-return trade off maximises equation (10), whereby the risk free

return is the return on the one month Treasury Bill on 31.12.98 of 4.47%. The historical distribution is

used to estimate the Value-at-Risk. The VaRs for the portfolio are given, the amount of borrowing or

lending required to meet the daily historical VaR at the 95%, and the final portfolio allocation such that the

VaR constraint is met.

Asset allocation to maximise risk-return trade-off

Empirical Stocks Bonds VaR Borrowing VaR*

95% 36% 64% 6.836 0 6.836

97.5% 51% 49% 10.016 -313.20 6.836

99% 34% 66% 11.400 -395.59 6.836

Optimal Portfolio to meet VaR constraint

Empirical Stocks Bonds Cash

95% 36.00% 64.00% 0.00%

97.5% 35.03% 33.65% 31.32%

99% 20.55% 39.89% 39.56%

22
Table 2

Optimal Asset Allocation under Normality and Student-t

The table gives the optimal allocation between Stocks and Bonds using data on the S&P 500 Composite

Returns Index and the 10- Year Datastream US Benchmark Government Bond Index over the period

January 1980 - December 1998. The risk-return trade off maximises equation (10), whereby the risk free

return is the return on the one month Treasury Bill on 31.12.98 of 4.47%. The normal distribution and the

Student-t with 5 degrees of freedom are used to estimate the Value-at-Risk. The VaRs for the portfolio are

given, the amount of borrowing or lending required to meet the daily historical VaR at the 95%, and the

final portfolio allocation such that the VaR constraint is met.

Asset allocation to maximise risk-return trade-off

Optimal Portfolio Normality Student-t

Stocks Bonds VaR Borrowing VaR Borrowing

95% 40% 60% 7.160 -44.43 6.771 9.45

97.5% 40% 60% 8.621 -203.82 8.766 -216.76

99% 40% 60% 10.32 -333.17 11.618 -406.80

Optimal Portfolio to meet VaR constraint

Normality Student-t

Normal Stocks Bonds Cash Stocks Bonds Cash

95% 38.22% 57.33% 4.44% 40.38% 60.57% -0.95%

97.5% 31.85% 47.77% 20.38% 31.33% 46.99% 21.68%

99% 26.67% 40.01% 33.32% 23.73% 35.59% 40.68%

23
Figure 1

Efficient VaR Frontier - 95% Student-t, Normal and Empirical VaR

The figure presents the risk return trade off for portfolios of Stocks and Bonds whereby risk is measured

by the VaR of the portfolio at the 95% confidence level. The returns and VaR estimates are obtained using

data on the S&P 500 Composite Returns Index and the 10- Year Datastream US Benchmark Government

Bond Index for the period January 1980 until December 1998. We present the efficient frontier for the

empirical distribution, the parametric normal approach and under the assumption of a Student-t

distribution with 5 degrees of freedom.

Efficient VaR Frontier: 95% Confidence

0.0007

0.0006

0.0005

0.0004
RETURN

0.0003

0.0002

0.0001

0
0 5 10 15 20 25
RISK

Empirical Normal Student-t(5)

Empirical CML Normal CML Student-t(5) CML

24
Figure 2

Efficient VaR Frontier - 99% Student-t, Normal and Empirical VaR

The figure presents the risk return trade off for portfolios of Stocks and Bonds whereby risk is measured

by the VaR of the portfolio at the 99% confidence level. The returns and VaR estimates are obtained using

data on the S&P 500 Composite Returns Index and the 10- Year Datastream US Benchmark Government

Bond Index for the period January 1980 until December 1998. We present the efficient frontier for the

empirical distribution, the parametric normal approach and under the assumption of a Student-t

distribution with 5 degrees of freedom.

Efficient VaR Frontier: 99% Confidence

0.0007

0.0006

0.0005

R 0.0004
E
T
U 0.0003

0.0002

0.0001

0
0 5 10 15 20 25
RISK

Empirical Normal Student-t(5)

25
Endnotes

1 See Jorion (1997) for a comprehensive introduction into Value-at-Risk methodology.

2 In practice these confidence levels range from 95% through 99%, whereby the Basle Committee
recommends 99%.
3 This is the VaR recommended by the Basle Committee for Banking Regulation used in establishing a
bank’s capital adequacy requirements.
4 A higher confidence level by definition will result in a higher Value-at-Risk.
5 In a similar manner specifying a confidence level below that used for the optimisation the risk manager
would want to take on additional risk by borrowing additional funds at the risk free rate, and going
short in the 3-month Treasury Bill.
6 On average in the sample the normal distribution significantly overestimates the risk-return trade off by
2.12%.
7 Maximising the Sharpe ration also generates an optimal asset allocation of 40% stocks and 60% bonds.
8 A market equilibrium model could however be derived under other assumptions, such as imposing the
assumption that all investors focus on the same confidence level for VaR.
9 See Huisman, Koedijk and Pownall (1998).
10 The smaller the number of degrees of freedom used to parameterise the student-t distribution the fatter
the tails of the distribution and the greater the severity of the difference between the normal
distribution. For consistency we use five degrees of freedom throughout the empirical analysis.
11 Of course the use of a fatter tailed distribution with more degrees of freedom will be better able to
capture the risk-return trade off as seen in the empirical VaR frontier.
12 See Longin and Solnik (1995) for empirical analysis into the instability of international covariance and
correlation matrices over time.
13 See Arzac and Bawa (1977) for the derivation of a market equilibrium model.
14 This result enables us to test the VaR approach to asset allocation when other distributional
assumptions are assumed.

26